Power-efficient data management for dynamic applications
Pol Marchal, José Ignacio Gómez, David Atienza, Stylianos Mamagkakis, F. Catthoor · Institution of Engineering and Technology eBooks · 2006
In dynamic multi-threaded applications dealing with dynamic data and tasks is crucial. The memory bandwidth is both an issue at the shared SDRAM memory as well as on the local memory layer. Modern multi-media applications contain multiple tasks and/or benefit from task parallelisation. However, tasks running on multiple processors can access the same memory in parallel. This causes access conflicts that delay the system and increase its energy cost. Since existing techniques also optimise the memory bandwidth inside a single task, they cannot cope with inter-task conflicts. A need thus exists for techniques that optimise the memory bandwidth across the tasks' boundaries. We have therefore introduced several task-ordering/data assignment techniques for both the local memory layer as the SDRAM. Multi-media applications consist of multiple tasks which are started/stopped at runtime due to user events. Also the tasks themselves have become data-dependent. We have shown that design-time nor runtime techniques can effectively deal with this dynamic behaviour. Therefore, we have introduced a novel scenario-based memory bandwidth approach. It combines the best of the design-time and runtime techniques. Finally, runtime memory management requires an efficient management of the free space. We have introduced a methodology which customises the DM managers for this purpose.